[Paper Review] A New Periocular Dataset Collected by Mobile Devices in Unconstrained Scenarios

Luiz A. Zanlorensi, Rayson Laroca+4 more|arXiv (Cornell University)|Nov 24, 2020

Biometric Identification and Security4 citations

TL;DR

This paper introduces the UFPR-Periocular dataset, the largest publicly available periocular dataset with 1,122 subjects and 196 mobile devices across three sessions under unconstrained conditions. Using a multi-task deep learning model based on MobileNetV2, the study achieves a rank-1 identification rate of 84.32% and an EER of 0.81% in the closed-world protocol, demonstrating the need for further research in mobile-based periocular recognition under real-world variability.

ABSTRACT

Recently, ocular biometrics in unconstrained environments using images obtained at visible wavelength have gained the researchers' attention, especially with images captured by mobile devices. Periocular recognition has been demonstrated to be an alternative when the iris trait is not available due to occlusions or low image resolution. However, the periocular trait does not have the high uniqueness presented in the iris trait. Thus, the use of datasets containing many subjects is essential to assess biometric systems' capacity to extract discriminating information from the periocular region. Also, to address the within-class variability caused by lighting and attributes in the periocular region, it is of paramount importance to use datasets with images of the same subject captured in distinct sessions. As the datasets available in the literature do not present all these factors, in this work, we present a new periocular dataset containing samples from 1,122 subjects, acquired in 3 sessions by 196 different mobile devices. The images were captured under unconstrained environments with just a single instruction to the participants: to place their eyes on a region of interest. We also performed an extensive benchmark with several Convolutional Neural Network (CNN) architectures and models that have been employed in state-of-the-art approaches based on Multi-class Classification, Multitask Learning, Pairwise Filters Network, and Siamese Network. The results achieved in the closed- and open-world protocol, considering the identification and verification tasks, show that this area still needs research and development.

Motivation & Objective

• To address the lack of large-scale, real-world periocular datasets collected via mobile devices under unconstrained conditions.
• To evaluate the performance of state-of-the-art deep learning models in periocular recognition under high intra-class variability caused by lighting, pose, and device differences.
• To investigate the impact of multi-task learning on improving discriminative feature extraction for identification and verification tasks.
• To provide a benchmark dataset and experimental setup for future research in mobile ocular biometrics.

Proposed method

• Collected 33,660 periocular images from 1,122 subjects across three sessions using 196 different mobile devices in unconstrained environments.
• Annotated eye corners manually and provided metadata including age range, gender, and device model for each image.
• Trained and evaluated multiple CNN architectures, including models for multi-class classification, multi-task learning, Siamese networks, and pairwise filters networks.
• Employed both closed-world and open-world protocols to evaluate identification and verification performance.
• Conducted an ablation study to assess the contribution of each task in the multi-task learning framework.
• Used data augmentation and attribute normalization techniques to mitigate errors from lighting, blur, occlusion, and eyeglasses.

Experimental results

Research questions

• RQ1 How does the performance of periocular recognition models vary across different deep learning architectures in unconstrained mobile scenarios?
• RQ2 What is the relative contribution of auxiliary tasks (e.g., gender, age, device model) in improving the discriminative power of periocular features?
• RQ3 How do lighting, occlusion, and image resolution affect verification accuracy in real-world mobile periocular recognition?
• RQ4 Can multi-task learning effectively reduce intra-class variability and improve generalization in mobile-based periocular recognition?
• RQ5 What is the performance gap between closed-world and open-world protocols in real-world periocular recognition?

Key findings

• The UFPR-Periocular dataset is the largest in terms of subject count (1,122) and number of unique mobile devices (196) in the literature for visible-spectrum periocular recognition.
• The multi-task learning model using MobileNetV2 achieved the best performance, with a rank-1 identification rate of 84.32% and an equal error rate (EER) of 0.81% in the closed-world protocol.
• In the open-world protocol, the same model achieved an EER of 2.81% with thresholds of 0.80 and 0.78, indicating robustness to unknown subjects.
• The ablation study revealed that device model identification was the most influential task, followed by age range, gender, and eye side classification.
• Subjective analysis showed that lighting, occlusion, and low image resolution were the primary causes of false verification decisions.
• Despite strong performance, the results indicate that mobile periocular recognition still requires significant improvement, especially in open-world and real-world conditions.